AVALON: A New Framework for Zero-Shot Indoor Navigation Without Mapping

Indoor service robots entering a building they have never set foot in, yet expected to complete tasks as efficiently as a user familiar with the place—this is a long-standing but rarely directly addressed scenario in robot navigation. Most existing approaches either rely on pre-built maps or require training for specific sites. A study from an RSS source proposes the AVALON framework, which attempts to break this limitation using "human-seeded hierarchical spatial memory."

The Core Problem: Human-Familiar, Robot-Unseen Spaces

The paper defines the target scenario as "human-familiar, robot-unseen"—the user knows the building layout and task constraints inside out, but the robot has never been to the site before. Under this setting, requiring the robot to first build a global SLAM map before navigating is neither economical nor realistic.

AVALON's deployment requirements are compressed to an extreme minimum: a floor plan, some place clues, and a single initial topological-odometry alignment. No prior on-site traversal, no site-specific model training. The robot does not need to build a global map itself—it relies on the existing topological structure to localize and execute navigation.

The key shift in this approach lies in redirecting the answer to "where does spatial knowledge come from"—from "the robot explores on its own" to "humans already know."

SLAM (Simultaneous Localization and Mapping) is a foundational technical paradigm in robot navigation: a robot builds a map while simultaneously using that map to estimate its own pose in an unknown environment. Traditional SLAM requires the robot to physically traverse the environment to produce a usable global map, and map quality depends heavily on sensor coverage and loop closure detection. This is precisely the cost AVALON aims to circumvent—in "human-familiar, robot-unseen" scenarios, the building already has human knowledge such as floor plans, so making the robot run SLAM from scratch is both time-consuming and wasteful. AVALON replaces robot-built SLAM with human-seeded topological memory, essentially trading "prior, semantic, sparse structure" for "posterior, geometric, dense maps." The cost is losing the precise local planning capability that continuous metric maps provide; the benefit is zero-shot deployment and faster task startup.

Topological node networks are the core abstraction of AVALON's spatial representation. Unlike occupancy grid maps that describe environments with continuous coordinates, topological maps retain only key locations (nodes) and their connectivity (edges)—like a subway map rather than a satellite photo. Nodes typically correspond to semantically salient locations such as rooms, corridor intersections, and doorways. This representation is naturally sparse, small in storage, and easy to align with human language descriptions, but it requires the robot to have "node verification" capability—when arriving somewhere, it must confirm that it has indeed reached the expected node, otherwise topological reasoning will cascade into errors. This requirement also explains why AVALON treats node verification as an independent layer.

How the Three-Layer Memory Architecture Works

AVALON's framework consists of three interlocking components: human-seeded hierarchical spatial memory, node-verified topological execution, and task-triggered, evidence-gated human-robot memory co-update.

The hierarchical spatial memory is the foundational layer. Humans provide floor plans and place clues, which are organized into a topological node network. The robot does not possess a global map but can know "which node it is near" within the topological structure. Visual depth point-goal navigation uses verified node arrival events for translational re-anchoring, thereby maintaining a sense of position without a robot-built SLAM map.

Node-verified topological execution is the middle layer. The robot does not blindly walk along topological edges but performs verification upon reaching a node, confirming that it has indeed arrived at the expected location. This verification mechanism provides trustworthy physical state evidence for subsequent memory updates.

Memory co-update is the most distinctive top-level design. Human operators are responsible for specifying task constraints that are visually unobservable and reporting offline environmental changes; the robot provides physical state evidence through active on-site verification. Task relevance determines which human-reported environmental changes need to be verified by the robot in the current task. Perceptual evidence serves as a gating condition that determines whether memory is updated—this ensures that a clear distinction always exists between "human-specified permissions" and "robot-observed traversability."

Translational re-anchoring is the key operation by which AVALON maintains a sense of position without a global SLAM map. Visual depth point-goal navigation typically relies on local odometry to estimate displacement relative to a target, but odometry accumulates drift over time, and after long tasks the deviation may become too large to correct. AVALON's approach is: whenever the robot confirms arrival at a verified topological node, it "re-anchors" the origin of the current odometry coordinate frame to that node's known coordinates—equivalent to continuously correcting the accumulated error of continuous odometry using discrete, reliable node arrival events. This is similar in spirit to loop closure detection in SLAM, but the correction source is not geometric feature matching—it is topological node arrival confirmation, making it more robust to visual appearance changes and better suited to scenarios where humans provide clues.

Evidence gating is an admission mechanism for memory writes. During human-robot co-update of memory, if human-reported environmental changes were written unconditionally, the robot might plan infeasible paths next time due to incorrect memory; conversely, if robot observations were written unconditionally, temporary obstacles might be misjudged as permanent changes. AVALON's solution is: memory is updated only when it is relevant to the current task and the robot has perceived corresponding physical evidence on-site. This effectively divides "semantic priors provided by humans" and "physical facts provided by robots" into two evidence tiers, and the gating mechanism commits persistent memory only when the two are consistent, thereby suppressing error accumulation during long-term operation.

Experimental Data: Efficiency and Success Rate Comparison

The research team conducted experiments on a Unitree Go2 quadruped robot. AVALON combined with the X-NavDP policy completed a 12.81-meter task from Intersection to Dining in 46.4 seconds, with all five runs successful (5/5), without any mapping throughout.

The comparison approach DimOS-Go2 took 78.5 seconds on the same task, succeeding four out of five times (4/5). AVALON's robot execution time was 40.9% less. Meanwhile, the approach using only the X-NavDP policy without the AVALON framework timed out on all five runs.

This set of data demonstrates two things: the topological framework itself brings significant efficiency gains, and policy-level improvements (X-NavDP) can hardly complete tasks independently in zero-shot indoor navigation scenarios without the support of topological memory.

X-NavDP is the navigation policy used in conjunction with AVALON in the experiments. Judging from the naming, it belongs to end-to-end navigation policies based on visual or depth input (DP may refer to Diffusion Policy or Depth-based Policy), responsible for outputting local actions based on current observations. Notably, experiments show that the pure X-NavDP approach timed out on all five runs, indicating that such policies lack long-term consistency in zero-shot unfamiliar environments—they can handle local obstacle avoidance and short-range goal approach but cannot maintain direction sense and goal consistency at the scale of tens of meters. The topological memory provided by AVALON precisely fills this layer: the policy handles "how to walk," while the topological layer handles "where to go" and "whether we've arrived." This layered division of labor suggests that combining end-to-end policies with explicit spatial memory may be more effective than simply scaling up policy models.

Why This Direction Deserves Attention

AVALON's value lies not only in a set of experimental data. Its proposed "human-robot memory co-update" mechanism touches on a more fundamental question: in robot deployment, how should human knowledge and machine perception divide labor?

Humans excel at providing high-level semantic constraints—"this passage closes after 3 PM," "that table cannot be moved"—information that visual systems cannot directly obtain. Robots excel at providing physical verification—"I did pass through this passage," "this location is currently occupied by an obstacle." AVALON uses an evidence gating mechanism to manage the two separately, avoiding confusion between human specifications and robot observations.

This design approach is particularly meaningful for indoor service robots that need to operate long-term in dynamic environments. Building layouts change, furniture moves, task constraints adjust—and approaches relying solely on pre-built maps struggle to keep up with these changes. AVALON incorporates "change" itself into the memory update process, allowing the system to gradually accumulate understanding of the new environment during operation.

Of course, the currently published experimental scale remains limited. Validation on a single task and a single robot platform is insufficient to demonstrate the framework's generalization capability in more complex scenarios. The scalability of the topological structure when facing large-area, multi-floor buildings, as well as the long-term stability of human-robot memory updates, are directions that need to be observed in follow-up work.

One-line Summary

AVALON uses human-seeded hierarchical topological memory to enable zero-shot navigation in unfamiliar buildings without mapping or site-specific training.

Paragraph Summary

AVALON targets the "human-familiar, robot-unseen" indoor navigation scenario, proposing to replace traditional robot-built SLAM maps with human-seeded hierarchical spatial memory. Deployment requires only a floor plan, place clues, and a single topological-odometry alignment—no prior traversal or site-specific training. The framework consists of three layers: hierarchical spatial memory, node-verified topological execution, and task-triggered, evidence-gated human-robot memory co-update. On a Unitree Go2, AVALON combined with X-NavDP completed a 12.81-meter task in just 46.4 seconds, with all five runs successful—40.9% less time than DimOS-Go2—while the pure policy approach timed out on all five runs. The core contribution of this work lies in separately managing human semantic priors and robot physical verification through an evidence gating mechanism, offering new ideas for long-term deployment in dynamic environments.